tree: 0d36d96275282434e5c82b451b98f8703c92ece5 [path history] [tgz]
  1. testdata/
  2. BUILD.bazel
  3. main.go
  4. main_test.go
  5. Makefile
  6. README.md
perf/nanostat/README.md

nanostat

nanostat compares statistics about nanobench results.

Installation

You need to have Go installed. Then run:

go get -u go.skia.org/infra/perf/nanostat

If you have $GOPATH or $GOBIN set then nanostat should appear in your path, otherwise it will be installed at $HOME/go/bin/nanostat.

Description

Each input file should be a JSON output file from nanobench.

Invoked on a pair of input files, nanostat computes statistics for each file and a column showing the percent change in mean from the first to the second file. Next to the percent change, nanostat shows the p-value and sample sizes from a test of the two distributions of nanobench results.

For example in the results below, ‘p’ equals 0.001 or 0.1%, and the analysis was done with 10 samples from the first file and 8 samples from the second file.

         old          new  delta     stats            name
  2.15 ±  5%   2.00 ±  2%   -7%   (p=0.001, n=10+ 8)  tabl_digg.skp

Small p-values indicate that the two distributions are significantly different. If the test indicates that there was no significant change between the two benchmarks (defined as p > alpha), nanostat displays a single ~ instead of the percent change.

Example

Suppose we collect benchmark results from running

out/Release/nanobench --config gl 8888 --outResultsFile old.json

Then make some changes to the code, recompile nanobench and run:

out/Release/nanobench --config gl 8888 --outResultsFile new.json

Then nanostat summarizes the differences between the old and new runs:

$ nanostat --iqrr old.json new.json
          old          new  delta         s            name
   0.78 ±  3%   0.72 ±  3%   -8%   (p=0.000, n=10+10)  desk_wowwiki.skp
   2.15 ±  5%   2.00 ±  2%   -7%   (p=0.001, n=10+ 8)  tabl_digg.skp
   3.08 ±  2%   2.96 ±  3%   -4%   (p=0.001, n= 9+10)  desk_facebook.skp
   0.71 ±  2%   0.69 ±  3%   -3%   (p=0.028, n= 9+10)  desk_ebay.skp
   4.59 ±  1%   4.46 ±  1%   -3%   (p=0.000, n=10+ 8)  desk_linkedin.skp
   1.40 ±  1%   1.39 ±  0%   -1%   (p=0.011, n= 9+ 9)  desk_css3gradients.skp
$

Usage

usage: nanostat [options] old.json new.json
options:
  -all
       If true then include insignificant changes in output.
  -alpha float
       Consider a change significant if p < α. Must be > 0. (default 0.05).
  -iqrr
       If true then remove outliers in the samples using the Interquartile Range Rule.
  -sort order
       Sort by order: [-]delta, [-]name (default "delta")
  -test string
       The type of test to do, 'utest' for Mann-Whitney U test, and 'ttest' for a Two Sample Welch T test. (default "utest")

To get help:

$ nanostat -h